README.md

metadata

license: apache-2.0
language:
  - en
pipeline_tag: text-classification
library_name: transformers
tags:
  - emotions
  - multi-task-learning
  - text-classification
  - sentiment-analysis
  - natural-language-processing
  - psychological-modeling
  - emotion-recognition
  - bert
  - deep-learning
  - emotion-embeddings
  - valence-arousal
  - emotion-ontology

🌌 EmotionVerse-2: Galactic Emotional Intelligence

Apache-2.0 BERT-base encoder Multi-Task (6 heads) Plutchik + Circumplex F1 Macro: 0.951 Coherence Index: 0.97 HF Transformers

EmotionVerse-2 unifies psychologically-grounded labeling (Plutchik) with a valence-arousal manifold, neural plasticity, and memory consolidation. It doesn’t just tag text—it resolves why it feels the way it does, across interlocking tasks that share a single encoder for maximal transfer.

✨ Revolutionary Emotional Paradigm

A single BERT encoder feeds six specialized heads—Primary, Secondary, Meta, Sentiment, Interaction, and Context—trained end-to-end with dynamic loss weighting. The result is emergent cross-task awareness: subtle shifts, mixed feelings, and psychologically consistent outputs.

• 🚀 Unified Multi-Task Engine. One encoder; six heads. Joint optimization learns shared affective features while preserving task-specific finesse.
• 🧬 Plutchik × Russell Fusion. Labels live in a combined categorical + circumplex space (valence, arousal, intensity), preserving family hierarchies and psychological distances.
• 🔗 Adaptive Loss Balancing. Dynamic task weights + focal components lift rare emotions without destabilizing common ones.
• 📊 Holistic Metrics. Per-task precision/recall/F1 plus a cross-task Emotion Coherence Index (valence-arousal similarity across heads).

View architecture diagram (SVG)

🏆 Grand Performance Showcase

The evaluation deliberately pits generalist GoEmotions models against a specialist trained on Plutchik’s eight core emotions. The “unfairness” is instructive: it exposes the cost of vocabulary mismatch and the power of a psychologically aligned label space.

🧠 Emotional Lexicons: Why Translation Fails

• Missing Concepts. GoEmotions lacks dedicated pathways for Plutchik categories like Trust or Anticipation as defined here.
• Forced Guesses. Inputs such as “I have complete faith in your plan” map to “Approval/Admiration,” not “Trust.” Close ≠ correct when precision matters.
• Semantic Misalignment. The learned geometry of the label space differs; a round peg in a square hole.

Model	F1 Macro
EMOTIONVERSE-2	0.951 🌟
GoEmotions-RoBERTa	0.024 🔻
GoEmotions-BERT	0.012 📉
DistilBERT-Emotion	0.007 👻

💎 Nearly-Perfect Diagonal (Confusion Matrix)

Diagonal dominance indicates surgical separation of Plutchik classes; off-diagonals remain low and psychologically adjacent when present.

Cross-task capability radar (SVG)

🔬 Architecture & Emotional Data Synthesis

Core Modules

• Neural Plasticity Gate. Reweights hidden states conditioned on intensity and local valence/arousal; supports context-sensitive adaptation.
• Amygdala-Inspired Intensifier. Saturating non-linearity emphasizes high-arousal signals without drowning low-arousal nuance.
• GRU Memory Consolidator. Injects sequential emotional context to stabilize predictions across turns or sentences.
• Circumplex Injector. Concatenates/merges valence-arousal (+ intensity) vectors into token-level or pooled embeddings.

Training Setup

• Backbone: BERT-base uncased.
• Optimization: AdamW, linear warmup/decay, mixed precision.
• Loss: Weighted sum across heads with dynamic task weights; focal term for rare labels; label smoothing for stability.
• Regularization: Dropout, stochastic depth on head adapters (if enabled), gradient clipping.

🚀 Psychologically-Informed Embeddings

Emotion labels are embedded vectors, not IDs. Selected dimensions:

• Valence/Arousal Anchors. First two dims: positivity/negativity and activation.
• Intensity Signal. Third dim calibrated from dataset statistics.
• Family Centroids. Pull toward Plutchik family centers to preserve conceptual proximity.
• Relationship Pull. Secondary affinities (e.g., Joy↔Anticipation) maintain a smooth manifold.

📊 Evaluation & Performance Horizons

• Primary Head: Accuracy, Top-2 recall.
• Secondary & Meta: Macro/Weighted F1.
• Interaction & Context: Precision/Recall for communicative style cues.
• Sentiment: Mixed-class accuracy with polarity overlap analysis.
• Emotion Coherence Index: Cross-head valence/arousal agreement.

🛠️ Usage & Integration

Quickstart (Transformers)

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

name = "ayjays132/EmotionVerse-2"
tok  = AutoTokenizer.from_pretrained(name)
mdl  = AutoModelForSequenceClassification.from_pretrained(name)
text = "I’m thrilled yet anxious about tomorrow’s launch."
batch = tok(text, return_tensors="pt", truncation=True, padding=True)
with torch.no_grad():
    out = mdl(**batch)

Example: Plutchik mapping & softmax

import torch.nn.functional as F

plutchik_labels = ["joy","trust","anticipation","surprise","anger","sadness","fear","disgust"]
plutchik_logits = out.logits[:, :len(plutchik_labels)]  # slice depends on export
probs = F.softmax(plutchik_logits, dim=-1)[0].tolist()
for lbl, p in sorted(zip(plutchik_labels, probs), key=lambda x: -x[1]):
    print(f"{lbl:12s} {p:.3f}")

Batch inference with coherence check

def coherence(valence_arousal_primary, valence_arousal_sentiment):
    # cosine similarity across valence/arousal heads
    a, b = valence_arousal_primary, valence_arousal_sentiment
    return (a @ b) / ((a.norm()+1e-8) * (b.norm()+1e-8))

📂 EmotionVerse Dataset: Cosmic Fuel

The EmotionVerse dataset anchors this model with 3K+ entries annotated across six dimensions, including meta-emotional and contextual narratives.

📚 Reproducibility Snippets

Minimal training loop sketch

from torch.optim import AdamW
from torch.cuda.amp import autocast, GradScaler

scaler = GradScaler()
optim  = AdamW(mdl.parameters(), lr=3e-5, weight_decay=0.01)
for step, batch in enumerate(train_loader):
    with autocast():
        out = mdl(**{k:v.to(mdl.device) for k,v in batch.items()})
        # combine task losses (weights dynamically updated)
        loss = out.loss  # if model returns weighted sum
    scaler.scale(loss).backward()
    scaler.step(optim); scaler.update(); optim.zero_grad()

Evaluation (macro F1) sketch

from sklearn.metrics import f1_score

  all_y, all_p = [], []

for batch in val_loader:
    with torch.no_grad():
        out = mdl(**{k:v.to(mdl.device) for k,v in batch.items()})
    logits = out.logits[:, :8]  # primary head slice
    preds  = logits.argmax(-1).cpu().tolist()
    all_p += preds; all_y += batch["labels"].tolist()
print("F1 Macro:", f1_score(all_y, all_p, average="macro"))

📜 Licensing

Released under the Apache 2.0 License. Use, modify, and ship with confidence.

🙏 Acknowledgements

Thanks to Hugging Face (transformers, datasets) and the research community advancing affective computing, psychological modeling, and trustworthy NLP.